Evaluating AI for data-driven energy modeling: Performance of large language models versus expert-engineered approaches

Accurately predicting residential energy consumption is essential for improving energy efficiency, informing policy, and supporting decarbonization efforts. However, traditional data-driven modeling requires substantial expertise in data preprocessing, feature engineering, and machine learning, creating barriers for many practitioners. This study evaluates whether large language models (LLMs) can lower these barriers by autonomously performing end-to-end predictive modeling using the U.S. Residential Energy Consumption Survey (RECS) 2020 dataset. Four frontier commercial LLMs, including GPT-5.1, Claude Opus 4.5, Grok 4.1, and Gemini 3 are evaluated under zero-shot and structured step-wise prompting conditions and compared against a human-developed model. Results show that LLMs can independently execute complex analytical workflows, including data cleaning, exploratory analysis, feature identification, and model training. Grok 4.1 produced the strongest zero-shot models, achieving R 2 values above 0.85, while Gemini 3 and GPT-5.1 required structured guidance to complete the pipeline. The human-developed model achieved the highest accuracy, demonstrating the continued value of expert-guided preprocessing and systematic optimization. Overall, the findings indicate that while LLMs do not yet match expert-engineered models in predictive precision, they offer substantial potential to democratize data-driven building-energy analysis by enabling non-experts to generate reliable, interpretable models through natural-language interaction. The study provides the first cross-vendor evaluation of commercial LLMs for residential energy prediction and, rather than a controlled head-to-head evaluated, offers an exploratory proof-of-concept that highlights their emerging role as accessible analytical assistants for non-expert users in building-energy research and practice.